ALO Knowledge Power Plant - A Report on Reading the Structure of Technology and Civilization

Artificial intelligence (AI) is rapidly becoming smarter. It has reached a level where it can answer questions, create videos, write code, and understand human language.

However, behind this progress lies a reality that is not easily seen. As AI grows, it requires more electricity, more semiconductors, and more cooling equipment.

Today's AI primarily relies on GPUs (Graphics Processing Units). GPUs excel at processing numerous calculations simultaneously, making them the core equipment for large-scale AI learning and inference.

However, their weaknesses are also evident. They consume significant power and generate a lot of heat. As models become larger, the costs for semiconductors and power also increase.

Neuromorphic computing is a concept that has emerged in response to these limitations.

While the name may sound unfamiliar, its meaning is simple: an attempt to create semiconductors that operate like the human brain. 'Neuro' refers to nerves, and 'morphic' means having a similar form.

The human brain does not continuously compute like a supercomputer. It sends signals only when necessary, compresses repeated experiences, and prepares for judgment with minimal energy.

This is why engineers are focusing on the brain. The human brain uses around 20 watts of energy to see, hear, remember, and control movement.

In conventional computers, data stored in memory is retrieved by the central processing unit or GPU, computed, and then stored again. This process involves repetitive data movement, consuming power and time. This is often referred to as the "Von Neumann bottleneck."

The power issues in the AI era are deeply connected not only to the volume of computation itself but also to the cost of data movement.

Neuromorphic computing aims to change this structure.

It involves placing memory and computation closer together, reacting only when a change occurs, and sending short signals only when a certain threshold is crossed, similar to neurons in the brain.

This is called a Spiking Neural Network (SNN).

Simply put, if existing AI is like a large factory that is always on, neuromorphic AI is closer to a smart factory that operates only when needed.

It does not compute if there are no changes, and it only processes the parts where changes have occurred. For this reason, it is gaining attention in fields where low power consumption and real-time responsiveness are crucial.

Typical application areas include robots, drones, autonomous vehicles, smartphones, and defense equipment.

These devices cannot always be connected to large servers. They must be able to react on-site immediately even if communication is lost, operate for extended periods on battery power, and move without delay.

An accident could occur while AI is waiting for a response from the cloud.

Semiconductor technology is still in its early stages. Neuromorphic chips have not yet replaced GPUs.

However, major companies have already released actual chips and begun pilot operations. Intel, a US semiconductor company, has unveiled its neuromorphic research chip Loihi 2 and conducted experiments in robot control, pathfinding, and real-time learning.

These are more akin to low-power, reactive AI platforms rather than chips for large AI servers.

IBM, a US technology company, has also proposed a structure that closely integrates memory and computation within the chip through its NorthPole chip.

The core idea is to perform computations close to where the data needs to be processed, rather than moving data long distances. This approach aims to simultaneously increase AI inference speed and power efficiency.

Korea is not yet a leading player in the commercialization of complete neuromorphic chips.

However, its strengths are clear. These include memory semiconductors, advanced packaging, foundry capabilities, and low-power process technology. High Bandwidth Memory (HBM), next-generation memory, image sensors, and automotive semiconductors, where Samsung Electronics and SK Hynix excel, can serve as important connection points in the neuromorphic era.

In particular, the core question for neuromorphic computing is not "How can we compute more?" but rather "How can we reduce data movement?"

The technology to reduce the distance between memory and computation aligns with areas where the Korean semiconductor industry already holds strengths. This is why we must look beyond the competition for finished AI chips and also consider opportunities in core components and the manufacturing ecosystem.

Of course, there are many hurdles to overcome.

The existing AI ecosystem is built around GPUs and a general-purpose software environment. Development tools, talent, and standards are also currently GPU-centric.

For neuromorphic computing to become widely adopted, not only hardware but also the software and algorithm ecosystem must grow in parallel.

Therefore, the near future is more about coexistence than replacement.

In data centers, GPUs will continue to handle large-scale AI, while neuromorphic-based chips are likely to expand their role in edge devices such as robots, drones, cars, and sensors.

Large models will be managed by central servers, and immediate responses will be handled by edge chips.

The statement that AI is becoming like the human brain is not an emotional expression. It is an industrial choice made to react faster with less power.

Civilization has grown on the foundation of computational power. However, the next civilization might be built not on the quantity of computation, but on the efficiency of response.